Deep-Learning-2017-Lecture5CNN.ppt
- 1. Lecture 5 Smaller Network: CNN ο¬ We know it is good to learn a small model. ο¬ From this fully connected model, do we really need all the edges? ο¬ Can some of these be shared?
- 2. Consider learning an image: ο¬Some patterns are much smaller than the whole image βbeakβ detector Can represent a small region with fewer parameters
- 3. Same pattern appears in different places: They can be compressed! What about training a lot of such βsmallβ detectors and each detector must βmove aroundβ. βupper-left beakβ detector βmiddle beakβ detector They can be compressed to the same parameters.
- 4. A convolutional layer A filter A CNN is a neural network with some convolutional layers (and some other layers). A convolutional layer has a number of filters that does convolutional operation. Beak detector
- 5. Convolution 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 -1 1 -1 -1 1 -1 -1 1 -1 Filter 2 β¦ β¦ These are the network parameters to be learned. Each filter detects a small pattern (3 x 3).
- 6. Convolution 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 3 -1 stride=1 Dot product
- 7. Convolution 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 3 -3 If stride=2
- 8. Convolution 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 3 -1 -3 -1 -3 1 0 -3 -3 -3 0 1 3 -2 -2 -1 stride=1
- 9. Convolution 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 3 -1 -3 -1 -3 1 0 -3 -3 -3 0 1 3 -2 -2 -1 -1 1 -1 -1 1 -1 -1 1 -1 Filter 2 -1 -1 -1 -1 -1 -1 -2 1 -1 -1 -2 1 -1 0 -4 3 Repeat this for each filter stride=1 Two 4 x 4 images Forming 2 x 4 x 4 matrix Feature Map
- 10. Color image: RGB 3 channels 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 -1 1 -1 -1 1 -1 -1 1 -1 Filter 2 1 -1 -1 -1 1 -1 -1 -1 1 1 -1 -1 -1 1 -1 -1 -1 1 -1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 Color image
- 11. 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 image convolution -1 1 -1 -1 1 -1 -1 1 -1 1 -1 -1 -1 1 -1 -1 -1 1 1 x 2 x β¦ β¦ 36 x β¦ β¦ 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 Convolution v.s. Fully Connected Fully- connected
- 12. 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 1 2 3 β¦ 8 9 β¦ 1 3 14 15 β¦ Only connect to 9 inputs, not fully connected 4: 10: 16 1 0 0 0 0 1 0 0 0 0 1 1 3 fewer parameters!
- 13. 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1 1: 2: 3: β¦ 7: 8: 9: β¦ 1 3: 14: 15: β¦ 4: 10: 16: 1 0 0 0 0 1 0 0 0 0 1 1 3 -1 Shared weights 6 x 6 image Fewer parameters Even fewer parameters
- 14. The whole CNN Fully Connected Feedforward network cat dog β¦β¦ Convolution Max Pooling Convolution Max Pooling Flattened Can repeat many times
- 15. Max Pooling 3 -1 -3 -1 -3 1 0 -3 -3 -3 0 1 3 -2 -2 -1 -1 1 -1 -1 1 -1 -1 1 -1 Filter 2 -1 -1 -1 -1 -1 -1 -2 1 -1 -1 -2 1 -1 0 -4 3 1 -1 -1 -1 1 -1 -1 -1 1 Filter 1
- 16. Why Pooling ο¬ Subsampling pixels will not change the object Subsampling bird bird We can subsample the pixels to make image smaller fewer parameters to characterize the image
- 17. A CNN compresses a fully connected network in two ways: ο¬Reducing number of connections ο¬Shared weights on the edges ο¬Max pooling further reduces the complexity
- 18. Max Pooling 1 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 0 0 1 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 6 x 6 image 3 0 1 3 -1 1 3 0 2 x 2 image Each filter is a channel New image but smaller Conv Max Pooling
- 19. The whole CNN Convolution Max Pooling Convolution Max Pooling Can repeat many times A new image The number of channels is the number of filters Smaller than the original image 3 0 1 3 -1 1 3 0
- 20. The whole CNN Fully Connected Feedforward network cat dog β¦β¦ Convolution Max Pooling Convolution Max Pooling Flattened A new image A new image
- 21. Flattening 3 0 1 3 -1 1 3 0 Flattened 3 0 1 3 -1 1 0 3 Fully Connected Feedforward network
- 22. Only modified the network structure and input format (vector -> 3-D tensor) CNN in Keras Convolution Max Pooling Convolution Max Pooling input 1 -1 -1 -1 1 -1 -1 -1 1 -1 1 -1 -1 1 -1 -1 1 -1 There are 25 3x3 filters. β¦ β¦ Input_shape = ( 28 , 28 , 1) 1: black/white, 3: RGB 28 x 28 pixels 3 -1 -3 1 3
- 23. Only modified the network structure and input format (vector -> 3-D array) CNN in Keras Convolution Max Pooling Convolution Max Pooling Input 1 x 28 x 28 25 x 26 x 26 25 x 13 x 13 50 x 11 x 11 50 x 5 x 5 How many parameters for each filter? How many parameters for each filter? 9 225= 25x9
- 24. Only modified the network structure and input format (vector -> 3-D array) CNN in Keras Convolution Max Pooling Convolution Max Pooling Input 1 x 28 x 28 25 x 26 x 26 25 x 13 x 13 50 x 11 x 11 50 x 5 x 5 Flattened 1250 Fully connected feedforward network Output
- 25. AlphaGo Neural Network (19 x 19 positions) Next move 19 x 19 matrix Black: 1 white: -1 none: 0 Fully-connected feedforward network can be used But CNN performs much better
- 26. AlphaGoβs policy network Note: AlphaGo does not use Max Pooling. The following is quotation from their Nature article:
- 27. CNN in speech recognition Time Frequency Spectrogram CNN Image The filters move in the frequency direction.
- 28. CNN in text classification Source of image: http://citeseerx.ist.psu.edu/viewdoc/downlo ad?doi=10.1.1.703.6858&rep=rep1&type=p df ?